Text-based emotion inference and empathetic response: Evaluating the capabilities of large language models relative to human counselors

In the face of the growing mental health crisis, researchers are increasingly exploring how large language models (LLMs) can be integrated into psychological support and counseling. Empathy—a core therapeutic factor in human counseling—plays an equally vital role in LLM-based interactions. This study systematically evaluates the empathetic capabilities of LLMs across two key dimensions of human empathy: emotion inference and empathetic response . Using 50 human counselors (20 novices and 30 experienced practitioners) as a benchmark, we evaluated the performance of three advanced LLMs—DeepSeek-R1, Qwen-Max, and GPT-4o—across text-based counseling scenarios. To address imbalanced data and enhance statistical rigor, we conducted linear mixed-effects modeling (LMM) and reported effect sizes with 95 % confidence intervals. Results revealed a significant crossover interaction: human counselors, particularly novices, demonstrated higher accuracy in inferring negative emotions, whereas LLMs exhibited higher accuracy in identifying positive emotions. For empathetic responses , both humans and LLMs exhibited higher response quality in positive contexts than in negative ones, with no significant group differences overall. These findings refine earlier interpretations by revealing distinct but complementary performance profiles between human counselors and LLMs. Collectively, the results suggest that advanced LLMs can demonstrate comparable capabilities to humans in specific emotional contexts and generate standardized, appropriate empathetic responses. By situating LLM performance within the framework of counseling theory and mixed-effects analysis, this study offers theoretical and practical insights into the evolving role of LLMs in mental health support and digital counseling practice.